Evidence Attributes as the Foundation of Governance Falsifiability

Artificial intelligence governance frameworks require evidence of compliance, yet systematically fail to specify what makes evidence trustworthy, observable, or re-auditable. This paper proposes Governance Falsifiability Infrastructure (GFI): an architecture that makes governance claims operational — observable, contradictable, and independently evidenced — without requiring the infrastructure itself to hold governance authority.

We identify two structural failures in current practice. First, governance evidence is typically produced by the system under assessment itself, making independent observation structurally absent from the moment evidence is created. Second, governance systems preserve outcomes but rarely preserve decision conditions, producing records that are complete but uninterpretable: a reassessor can reconstruct what was decided, but not whether a different decision would have been defensible under the standard then in force. We term this the loss of decision counterfactuality.

From these failures we derive an Observability Principle: a governance claim is assessable if and only if the observation channels required to falsify or corroborate it were active at the time the claim became operational. NOT ASSESSABLE is not a failure mode — it is a structural property of governance records that distinguishes absence of evidence from evidence of absence.

The architecture addresses both failures through three contributions. First, a formal Evidence Attribute Schema — a controlled vocabulary of atomic and derived observable facts, from which provenance categories (A0–A5) are derived using explicit, auditable rules, without normative judgment. Second, a Decision Baseline — a timestamped, sealed artifact capturing the governance context active at the moment of decision, independently of its outcome, enabling re-auditability as a formal property of the governance record. Third, a Consistency Engine operating under a No-Interpretation Principle: producing CONFIRMED, CONTRADICTED, or NOT ASSESSABLE results for each governance claim, without assigning governance meaning to those results.

Admissibility thresholds — determinations of whether evidence is sufficient — are externalized into Policy Packs: declarative, versioned, attributed specifications of required attributes and admissible provenance classes. Policy Packs behave as declarative queries over an Attribute Catalog, reducing the mapping between m evidence sources and n governance frameworks from m × n custom adapters to m + n independent components.

The architecture enforces a single principle throughout: no governance meaning below the Policy Pack layer. Every judgment is explicit, versioned, attributed to an authority, and contestable. The system makes governance claims observable, contradictable, and independently evidenced. It does not decide whether they are sufficient. Governance authority remains with the actor empowered to interpret the results.

Keywords: AI governance, evidence attributes, provenance classification, governance falsifiability, decision baseline, consistency engine, admissibility, observability, Policy Pack, EU AI Act